AWS Cost Optimization:
20 Strategies That Save Millions

Compute Savings Plans provide discounts of up to 66% on EC2, Lambda, and Fargate in exchange for a consistent $/hour spend commitment across any instance family, size, region, OS, and tenancy. This is the most flexible AWS commitment product and is the default recommendation for most EC2 workloads. Analyse 3–6 months of usage data using AWS Cost Explorer's Savings Plan recommendations tool to identify the optimal commitment amount. Start conservatively — commit to cover 60–65% of your baseline, not 100%, to leave room for workload changes. Commitment period: 1 year (66% discount) or 3 years (additional ~5% on top).

EC2 Instance Savings Plans offer up to 72% discount in exchange for a commitment to a specific instance family in a specific region. They are less flexible than Compute Savings Plans (you're committing to a family, not just a dollar amount) but deliver higher discounts for workloads with stable instance family requirements. Ideal for production application tiers with predictable instance family usage (e.g., m6i, c6g). Can be combined with Compute Savings Plans — use EC2 Instance SP for the most stable workloads, Compute SP for the remainder.

AWS Savings Plans do not apply to RDS, ElastiCache, OpenSearch, or Redshift. For these database services, Reserved Instances are the primary commitment discount mechanism. RDS RIs offer 30–60% savings over on-demand for 1 or 3-year terms. For production databases that run continuously, RIs are strongly recommended — database workloads are typically the most stable and predictable in an AWS estate. Convertible RIs for RDS allow instance type changes at renewal, providing flexibility if your database tier evolves.

Purchasing commitments is only half the work — unused commitments represent wasted spend. Monitor RI utilisation rate (target: above 90%) and Savings Plan utilisation weekly using AWS Cost Explorer. If utilisation drops, investigate whether workloads have been decommissioned or right-sized below your commitment. Unused Standard RIs can be listed for sale on the AWS Marketplace; Convertible RIs can be exchanged for different configurations. Build a monthly commitment review process into your FinOps operating model.

The most cost-efficient AWS compute strategy layers three pricing models: Savings Plans covering the steady-state baseline, Spot instances for variable and batch workloads, and on-demand for peak traffic spikes. This tiered approach can achieve effective average compute costs of $0.02–$0.04/vCPU-hour for general workloads — 70–85% below pure on-demand pricing. Build this layering into your Auto Scaling Groups using mixed instance policies to automatically blend pricing models based on workload characteristics.

AWS Compute Optimizer uses machine learning to analyse CloudWatch metrics and identify EC2 instances, Auto Scaling Groups, EBS volumes, and Lambda functions that are over-provisioned. The service is free (with optional enhanced infrastructure metrics add-on) and typically identifies 15–30% of instances as candidates for downsizing. Implement rightsizing in waves — start with non-production environments, then progressively address production workloads with adequate monitoring and rollback plans. Rightsizing before purchasing Savings Plans is critical: commit to the optimised baseline, not the over-provisioned one.

AWS Graviton3 and Graviton4 processors deliver 20–40% better price-performance than equivalent x86 instance types at the same or lower price. Graviton instances (m7g, c7g, r7g families) are available across EC2, RDS, ElastiCache, Lambda, and container services. Workloads that migrate most easily include: containerised applications, Java services on JVM, Python applications, and Go services. The main compatibility consideration is architecture-specific native binaries — most modern cloud-native applications have minimal friction. Running Graviton also unlocks lower Savings Plan rates for the same discount tier.

Spot instances offer 60–90% savings vs on-demand at the cost of potential 2-minute interruption notice. Suitable workloads include: CI/CD build pipelines, ML training jobs, data processing and ETL, rendering, game servers, and test environments. Use Spot Fleet or EC2 Auto Scaling Groups with multiple instance types and Availability Zones to minimise interruption risk — with 10–15 instance type options configured, interruption rates below 5% are typical. See our Spot Instance Strategy guide for interruption handling patterns.

Dev, staging, and QA environments that mirror production in size but are only used during business hours represent a significant waste opportunity. Implementing AWS Instance Scheduler or Lambda-based scheduling to stop non-production EC2 and RDS instances outside hours (e.g., 6pm–8am weekdays, all weekend) typically reduces non-production compute costs by 60–70%. This requires coordination with development teams to accommodate scheduled tasks or overnight test runs, but the savings — often 10–18% of total AWS spend — justify the process change.

For event-driven, low-frequency, or highly variable workloads, Lambda's consumption pricing model (per-millisecond billing, no idle cost) can dramatically reduce costs compared to always-on EC2 instances. A workload running 10 minutes per hour on a dedicated EC2 instance incurs full instance cost; on Lambda, you pay only for execution time. Lambda is not a universal replacement for EC2 — cold starts, execution time limits, and memory constraints apply — but reviewing workloads currently on EC2 that have bursty or infrequent invocation patterns can identify significant optimisation opportunities. See our Serverless Cost Optimisation guide.

S3 storage costs accumulate silently. S3 Standard at $0.023/GB/month becomes significant at petabyte scale — many enterprises discover they're paying millions annually for infrequently accessed data in hot storage tiers. S3 Intelligent-Tiering automatically moves objects to cheaper tiers based on access patterns with no retrieval fees (small monitoring fee per object). For data with known access patterns, S3 Lifecycle Policies can transition objects to S3 Standard-IA after 30 days, S3 Glacier Instant Retrieval after 90 days, and S3 Glacier Deep Archive after 180 days — reducing storage costs by 40–92% for cold data. See our Cloud Storage Optimisation guide.

Unattached EBS volumes — orphaned when EC2 instances are terminated — are a silent cost leak. Audit for volumes not attached to any running instance and either snapshot-then-delete or terminate them. Additionally, review volume types: many workloads run on gp2 volumes that could migrate to gp3 at 20% lower cost with better baseline performance. Compute Optimizer provides EBS right-sizing recommendations; reviewing these quarterly and acting on them is a consistent source of incremental savings.

AWS charges $0.09/GB for data transfer from EC2 to the internet in most regions, and $0.01–$0.02/GB for cross-AZ traffic. For data-intensive workloads, egress can represent 10–20% of total AWS spend. Strategies to reduce egress: consolidate workloads in the same AZ where latency permits (eliminating cross-AZ costs); use CloudFront to cache and serve static assets instead of direct S3-to-internet delivery; architect applications to minimise cross-region data movement; negotiate reduced egress rates as part of EDP discussions for high-egress workloads. See our Cloud Egress Costs guide.

Application Load Balancers ($0.008/LCU-hour plus hourly charges), NAT Gateways ($0.045/GB + hourly), and Elastic IP addresses ($0.005/hour when not attached) accumulate costs even with low or zero traffic. Many enterprises have load balancers provisioned for test environments that were never decommissioned. A quarterly audit of LBs, NAT Gateways, and EIPs against actual traffic metrics in CloudWatch identifies candidates for consolidation or removal. NAT Gateway optimisation — using VPC endpoints for S3 and DynamoDB access instead of routing through NAT — can reduce data processing costs significantly for high-volume workloads.

RDS instances are often over-provisioned for peak load and run at 10–30% average CPU utilisation. AWS Compute Optimizer and Performance Insights provide rightsizing recommendations for RDS — review quarterly and apply downsizing after adequate testing. Multi-AZ deployments double RDS instance costs in exchange for automatic failover — appropriate for production databases, but unnecessary for dev, test, and analytics databases. Auditing Multi-AZ deployments across all environments and disabling it where not required is a quick win with immediate cost impact.

Aurora Serverless v2 scales database capacity automatically from a configurable minimum ACU (Aurora Capacity Unit) to maximum, billing only for capacity used per second. For databases with highly variable load profiles — such as development databases, reporting databases with scheduled queries, or SaaS applications with uneven tenant activity — Aurora Serverless v2 can reduce costs by 40–80% compared to provisioned Aurora with equivalent peak capacity. Minimum ACU configuration should be set carefully to avoid cold start latency for latency-sensitive production workloads.

Amazon Redshift provisioned clusters that are idle (no queries running) still incur full hourly charges. For data warehouses used for periodic reporting (daily, weekly) rather than continuous analytics, pausing the cluster between query periods and resuming before scheduled runs can reduce Redshift costs by 50–80%. Redshift Serverless — which charges per compute-second of actual query execution — is an alternative for sporadic or unpredictable query workloads. Evaluate the break-even point: if your Redshift cluster runs queries less than 30% of the time, Serverless pricing is likely more economical.

The AWS EDP provides negotiated discounts (typically 10–25% depending on commitment size) across most AWS services in exchange for a multi-year minimum spend commitment. EDP becomes viable at approximately $1M annual AWS spend. Key negotiation variables include: discount rate by service category, eligible services (some managed services are excluded from EDP discounts), shortfall provisions (what happens if you underspend), ramp schedules (graduated year-1 commits), support credit bundles, and Marketplace spend allowances. Engaging an independent AWS negotiation advisor who can benchmark recent deal terms is strongly recommended — see our cloud negotiation firm rankings for vetted advisors.

AWS invests heavily in win-on-premises migrations and competitive cloud migrations (from Azure or GCP). If you have material on-premises workloads targeted for AWS migration, or existing workloads on competitor clouds, this creates significant negotiation leverage — AWS may offer migration credits, accelerated EDP ramps, or deeper discounts to secure the incremental spend. Quantify the migration pipeline value concretely (projected annual spend, timeline) and use it explicitly in EDP negotiations. Enterprises that secured 20%+ EDP discounts frequently attribute the additional headroom to credible migration commitments.

Organisations that implement cost governance before scaling AWS adoption avoid the expensive remediation work of cleaning up a sprawling, untagged, unmanaged estate. AWS Control Tower with Service Control Policies, mandatory tagging enforcement via AWS Config Rules, budget alerts via AWS Budgets, and a defined account vending model with cost allocation tags built-in from day one create a foundation that scales sustainably. The cost of retroactive governance — auditing thousands of resources, establishing allocation models for untagged spend, negotiating with engineering teams to add tags to production resources — far exceeds the upfront investment in a clean landing zone architecture. See our Cloud Cost Governance guide.